Hitherto, various methods for detecting the position of a pupil from an eye image are proposed, and for example, a method of binarizing image data of the eye image (hereinafter, abbreviated as “eye image data”) and detecting a circular area in an area of low-luminance level, and a method of calculating a contour integral of an image luminance I (x, y) with respect to an arc of a circle having a radius r and center coordinates (x0, y0) and calculating a partial derivative of the calculated amount relating to r in association with increase in the radius r are known. The structure in the aforementioned related art is disclosed, for example, in JP-T-8-504979. Several methods of increasing detection accuracy by eliminating the effects of eyelash or sunlight are also disclosed. The structure in the aforementioned related art is disclosed, for example, in JP-A-2002-119477.
In order to detect the pupil with high degree of accuracy using these methods, it is necessary to process a huge amount of image data at high-speed, and hence it is difficult to process the image data of the eye image on real time basis even though a large CPU having a high processing capability or a bulk memory is used in the status quo. Also, when the processing amount of the CPU is reduced to a degree which enables real time processing of the image data, there may arise a problem such that the detection accuracy is lowered.